Build creative agent

Build a Creative Agent

A creative agent accepts assets from buyers, stores or transforms them, and returns serving artifacts (tags, rendered manifests, previews). The fastest path to a passing agent is to fork a worked adapter that matches your archetype. This skill tells you which one and what cross-cutting rules apply.

Pick your fork target

Three creative archetypes. Two have dedicated worked adapters; the third (creative-generative) reuses the template adapter and applies brief-to-creative deltas.

Specialism	Archetype	Fork this	Mock upstream	Storyboard
creative-ad-server	Stateful library, pricing + billing (Innovid, Flashtalking, CM360, GAM-creative)	hello_creative_adapter_ad_server.ts	npx adcp mock-server creative-ad-server	creative_ad_server
creative-template	Stateless transform from inline manifest (Celtra, AudioStack, ElevenLabs, Resemble)	hello_creative_adapter_template.ts	npx adcp mock-server creative-template	creative_template
creative-generative	Brief-to-creative generation (standalone generative platform; not coupled with selling)	Same — fork the template adapter and apply the generative deltas below	Same	creative_generative

The interaction_model in each specialism's index.yaml is the forcing function: stateful_ad_server, stateless_transform, stateless_generate. Decide which one matches your business, then fork the adapter for that row.

Coupling generation with selling inventory? That's the AI-ad-network shape (e.g., generative DSPs that sell programmatic inventory and generate the creative). See skills/build-generative-seller-agent/ — combine sales-non-guaranteed + creative-generative.

For exact response shapes, error codes, and optional fields, docs/llms.txt is the canonical reference. The fork target stays in sync with the spec because PR #1394's three-gate contract fails CI when it drifts.

When to use this skill

• User wants to build an ad server, creative management platform, or audio creative agent
• User mentions sync_creatives, build_creative, list_creative_formats, or VAST/tag generation

Not this skill:

• Brief-to-creative generation coupled with selling inventory (AI ad network) → skills/build-generative-seller-agent/
• Publisher creative service that pairs with selling — usually a build-seller-agent adopter that omits a creative-* specialism claim

Cross-cutting rules

Every creative agent hits the cross-cutting rules in ../cross-cutting.md. The high-traffic ones for creative agents (deep-linked to the rule):

• idempotency_key on sync_creatives, build_creative, report_usage
• Authentication — serve({ authenticate }) baseline
• Webhooks — creative_review.${creative_id} and report_usage.${report_batch_id} operation_ids must be stable across retries

Two creative-specific notes on top of those:

Webhooks for async review pipelines

Creative review is naturally async. sync_creatives may return pending_review with a task envelope; your review pipeline emits creative_review completion webhooks when state transitions. build_creative for the ad-server archetype emits report_usage completion webhooks. Use ctx.emitWebhook per the pattern in ../cross-cutting.md § Webhooks.

previewCreative and listCreativeFormats are no-account tools

Both run with NoAccountCtx<TCtxMeta> — they don't carry an authenticated account. The fork target's resolver synthesizes a default-listing network so ctx.account resolves cleanly inside the no-account handler (migration recipe #11). If you fork from scratch you'll trip this; let the hello adapter do it.

preview_creative should return urlRender or bothRender, not htmlRender alone. The creative-template storyboard asserts previews[0].renders[0].preview_url is renderable. If your platform can host a preview URL, use urlRender. If you only have inline HTML, use bothRender (emit both preview_url and preview_html).

Preview URLs must resolve from durable storage in production; do not back MCPUI preview links with pod-local in-memory maps except in local demos. If you fork hello_creative_adapter_template.ts, replace the mock upstream preview_url plumbing with a route that reads preview bytes and session metadata from shared storage before shipping. See docs/guides/PREVIEW-ASSET-DURABILITY.md.

Specialism deltas at a glance

The fork targets cover the baseline + specialism deltas. Quick reference for what each archetype needs that the others don't:

creative-ad-server — stateful library (POST/GET/PATCH /v1/creatives), tag generation via macro substitution against a stored snippet, per-creative pricing (include_pricing=true on list_creatives, pricing_option_id echoed by build_creative, report_usage closes the loop), getCreativeDelivery with multi-id pass-through and required top-level currency + reporting_period. Output formats from list_creative_formats are serving-tag formats (VAST 4.2, display tag HTML, native JSON), not input visual formats.

creative-template — stateless: build from inline creative_manifest in the request, no ctx.store lookup. Multi-format path: target_format_ids (plural) returns { creative_manifests: [...] } (auto-wrapped to buildCreativeMultiResponse); single returns { creative_manifest: ... }. Formats declare variables[] the template substitutes.

Audio templates (TTS / mix / master, e.g. AudioStack / ElevenLabs / Resemble) follow the same shape with two deltas: (a) format declares parameterizedRender({ role: 'primary' }) because audio has no width/height — encode duration/codec/bitrate in accepts_parameters, not in render entries; (b) the upstream pipeline is naturally minutes-long (voice-over → mix → master), so the adapter returns a task envelope and emits creative_review webhooks on completion.

Audio adopters extend hello_creative_adapter_template.ts with four deltas:

1. Add audio_url?: string to UpstreamRender.output and extend output_kind to include 'audio_url'
2. Add an audio_url arm to outputSlot(t) returning FormatAsset.audio({ asset_id: 'serving_tag', required: false }) — declares the output slot in the format so build_creative's response key matches a declared asset_id per creative-manifest.json:14.
3. Add an audio-output branch to projectRenderToManifest:

   } else if (out.audio_url) {
     assets['serving_tag'] = audioAsset({ url: out.audio_url });
   }
   

   The audioAsset builder injects the asset_type: 'audio' discriminator the creative-manifest oneOf requires.
4. Seed an audio template in your upstream platform — see tpl_audiostack_spot_30s_v1 in src/lib/mock-server/creative-template/seed-data.ts for the worked reference.

The creative_template storyboard includes an optional audio_build phase. Enable it with a test kit that sets supports_audio_formats: true once your list_creative_formats response declares an audio format; the hello adapter CI gate asserts both the discovered audio format and the seeded AudioStack-style build output.

creative-generative

Forking hello_creative_adapter_template.ts for a creative-generative agent? Apply these deltas — the template adapter already handles the stateless build_creative flow, manifest projection, and async render polling. Generative just changes the input semantics (brief, not template inputs) and adds brand resolution.

• build_creative accepts message + brand.domain — no template_id, no slot inputs. Treat message as the creative brief and brand.domain as the brand identity.
• Honor quality: draft|production — drafts are cheap/fast (lower-quality model, no human review); production goes through full safety + brand checks.
• Refinement loop — buyers re-send the prior creative_manifest plus a new message to iterate. Detect "iterating on this manifest" by manifest-id presence and route to your refine endpoint instead of generate.
• Brand resolution via sync_accounts — buyers sync an account with brand.domain; treat that domain as resolvable for subsequent build_creative calls. Don't hardcode a brand allowlist; storyboards use fictional .example TLD brands.
• Response shape stays the same — { creative_manifest: { format_id, assets } }, not { creative_id, status, preview_url } (those are sync_creatives fields).
• Delete the template-input slot validation in the template adapter's build_creative handler — generative agents accept free-form briefs and can't validate slot presence/types up front.

The creative_generative storyboard exercises the same async render contract as creative_template — build_creative returns a render id, the buyer polls until status flips to complete. The template adapter's polling pattern transfers verbatim.

If your platform also sells inventory (AI ad network), claim both sales-non-guaranteed + creative-generative and see ../build-generative-seller-agent/ for the combined wiring.

Validate locally

# Run the fork-matrix gates for both creative archetypes
npm run compliance:fork-matrix -- --test-name-pattern="hello-creative-adapter"

# Or validate your forked agent directly against its storyboard
adcp storyboard run http://127.0.0.1:3002/mcp creative_ad_server \
  --bearer "$ADCP_AUTH_TOKEN" --include-bundles --json

Each gate is the three-gate contract from docs/guides/EXAMPLE-TEST-CONTRACT.md: tsc strict / storyboard zero-failures / upstream façade. Adopters who fork extend the test file with their own adapter path and expectedRoutes.

For deeper validation (fuzz, request-signing grading, custom invariants): docs/guides/VALIDATE-YOUR-AGENT.md.

Common shape gotchas

BuildCreativeReturn has 4 valid shapes (framework auto-wraps the bare manifest). VASTAsset requires an embedded delivery_type discriminator. Format.renders[] is a oneOf — dimensions (width + height) OR parameters_from_format_id: true, never both, never neither. See ../SHAPE-GOTCHAS.md — schema validators catch these at runtime; type checkers don't.

Migration notes

• 6.6 → 6.7: defineCreativeAdServerPlatform family drops req: unknown casts; Format.renders[] codegen is closed-shape (#1325) so typed render builders compose under strict tsc; NoAccountCtx<TCtxMeta> narrows the no-account tools. See docs/migration-6.6-to-6.7.md.
• 4.x → 5.x: docs/migration-4.x-to-5.x.md.